# Your pipelines. Your runners. One control plane. > Hive evaluates real pipeline code, schedules its DAG across the hardware you connect, keeps expensive inputs warm and exposes every run to people, scripts and agents through the same API. Source: https://hiveci.io/platform Language: en Alternate: https://hiveci.io/de/platform The hosted control plane is operated in Germany. Your runners execute wherever you place them. ```starlark load("@hive//ci.star", "standard_ci") load("@hive//checks.star", "checks") IS_PR = ctx.event == "pull_request" # computed in the entrypoint, every run ci_img = checks.ci_deps_image() fixer = autofix_step(target = "test", when = IS_PR) # dormant, PR-only hive.pipeline(**standard_ci( gate = checks.quality_gate(changed_files = ctx.changed_files, image_from = "ci-deps"), tests = backend_shards + ui_shards, extra = [rust_step, frontend_step], pr_only = [test_mobile, test_cli], # each built with `when = … and IS_PR` autofix = [fixer], images = [ci_img], )) ``` A complete pipeline file: the gate, the test shards, the extras, and one dormant fixer that wakes only on a pull request. ## The real gate, on your uncommitted tree, without a commit. hive check uploads a snapshot of your working tree and runs one step and its dependencies on the fleet. It evaluates the pipeline from the snapshot’s own .hive tree, so a broken Starlark edit fails here before it is pushed anywhere. The run is labelled local: it never notifies, never triggers a fixer and never collides with a real run. A red step prints its error and log tail in the terminal. It needs a reachable server, a token and a registered project — it is the gate, run early, not a local emulation of one. ```shell $ hive check --step test-2 3 changed file(s) vs base running only: test-2 (+ deps) run #4821 started on your working tree (base 9f3c2ab) https://app.hiveci.io/runs/… run #4821 failed ── test-2 (failed) ── error: exit status 1 log tail: FAILED orders/tests/test_total.py::test_order_total_includes_shipping 1 failed, 1261 passed in 298.11s ``` ## A build system worth keeping even without an agent. The agent layer is the differentiator. The scheduler, execution and data plane underneath are the reason its fixes can be verified. - **Definition: Pipelines as real code** — Starlark gives you functions, loops and reusable modules while evaluation stays deterministic and reviewable beside the repository. - **Placement: Run where the work fits** — Connect Linux, macOS, Windows or Kubernetes capacity. Labels, constraints, architectures, queues and caps shape placement. - **GPU jobs: GPU work lands where the node can run it** — Connect your own GPU machines; the agent reports model, count and driver. A step declaring hive.resources(gpu=N) lands only on a node with that many free GPUs and, when it names a minimum CUDA version, a driver that supports it. GPU-only nodes take nothing else, and with no GPU free the task waits and says why. - **Completion webhooks: Your application hears when a run ends** — Subscribe a URL per project or pipeline to succeeded, failed and cancelled runs. Each delivery is HMAC-SHA256 signed, carries the run and its artifact list, and is retried with backoff for up to a day; a stable delivery id lets the receiver drop duplicates. - **Warm path: Start from warm inputs** — Fingerprint-keyed images, registry caches and template test databases reduce repeated setup across nodes and runs. - **Evidence: Tests and timing become platform data** — Parsed reports, flaky-test history, step duration, queue wait, artifacts and logs accumulate without separate instrumentation. - **Interfaces: One answer across web, CLI, API and MCP** — Runs, logs, tests, fleet state and actions are first-class calls, so operators and agents do not scrape the UI. - **Isolation: Tenant and lane boundaries stay explicit** — Identity, scoped tokens, queue policy, resource limits and guarded actions are enforced by the control plane around each workload. ## Bring hardware first. Rent bursts when it helps. The hosted product keeps the control plane out of your operations budget while leaving compute placement under your control. - **Your runners: Connect machines you already own.** — A pull-based agent joins a machine or VM to the scheduler. Compute on those runners is not billed by Hive. - **Booster capacity: Spill bursts onto rented hardware.** — CPU-minute packs fund hosted execution when your own capacity is busy or the task needs a lane you do not operate. - **Hosted control plane: We operate the coordination layer.** — Projects, pipelines, runs, logs, artifacts, agents and policy stay available without asking your team to run Hive itself. ## Batteries included for the stack you already have. Not a plugin marketplace to assemble. These are first-class in the pipeline language, with the caches, the report parsing and the integrations already done. ### Languages: Lint and test, already wired One line per lane in your pipeline, and the runner, the cache key and the report format come with it — not a shell script you maintain. Python · pytest · ruff · uv · TypeScript · React · Vite · vitest · ESLint · pnpm · Rust · cargo · nextest · clippy · Go · Playwright ### Results: Failures as data, not scrollback Test output is parsed server-side, so a red run names the failing tests instead of handing you a log. Go streams natively; pytest and vitest stream through a one-line marker, so the pass/fail bar moves while the suite is still running. Flake and duration trends accumulate per test with nothing for you to instrument. go test -json · JUnit XML · live markers · coverage.xml · ESLint JSON · ruff JSON ### Build & ship: From Dockerfile to rolled out Images build on BuildKit against a registry-backed layer cache shared across nodes, then deploy with the rollout actually verified — including the GitOps commit that makes it survive a reconcile. Docker · OCI · BuildKit · Kubernetes · ArgoCD · CycloneDX · Trivy ### Where you work: It lives in your tools Checks, statuses and PRs land in GitHub. Tickets move in Linear — an agent can take one and open the PR for it. And Hive exposes its own tools over MCP, so your agents can drive it directly. GitHub · GitHub App · Linear · MCP · Starlark · PostgreSQL · S3 ### Models: Your key, your provider Agents authenticate as you — a Codex or Claude subscription, an OpenRouter key routing to whatever model you like, or Cursor. Spend is reported per session against your own account, at your provider’s rates. OpenAI Codex · OpenRouter · Anthropic Claude · Cursor · GLM Not on the list? Every step is a container running a command you choose, so anything that builds in CI builds here — Java, C++, Elixir, whatever you have. The names above are the ones Hive is optimised for, not the ones it is limited to. ## Integrations that exist in the code, listed by name. | | | |---|---| | GitHub | A GitHub App: per-installation tokens, webhooks, commit statuses, per-step check runs, pull-request comments and native auto-merge. | | Linear | A ticket lane for agents — claim, watch, comment, move state — plus the Factory’s own client. | | Sentry | Both directions: Hive reports its own errors, brokers Sentry reads to agents, and files deduplicated tickets from issues. | | Grafana | A query tool over the fleet metrics sink, for people and for agents. | | Kubernetes | Managed clusters as the execution substrate, plus a read and action lane — logs, describe, events, scale, restart, rollback — with changes proposed as GitOps pull requests. | | Registries and S3 | Container registries modelled as data with shared layer caching; artifacts, dependency caches and log archives in S3-compatible object storage. | | Notifications | ntfy and Discord webhooks, delivered through a durable outbox. | ## Move from orchestration to a control loop that can act on its own evidence. - **Turn eligible red checks into candidates** — The run wakes a scoped agent; the resulting branch gate remains authoritative. - **Operate sessions across the same fleet** — Workstation, cloud and Factory agents share the workspace around the pipeline. ## Bring one repository. See the whole system work. Start on your own runner, keep your existing CI beside it, and watch Hive turn the next eligible red check into a reviewable fix.